Gemini 3.8 Flash TTS Voice Remixing Is Live—Tune Timbre With Plain Language

Google’s public account (@Google) posted that Voice Remixing for Gemini 3.8 Flash TTS is now live in Google AI Studio and the Gemini API (https://x.com/Google/status/2118027058794148099). The post—timestamped 2026-11-04T17:06:16Z, or 2026-11-05 01:06:16 Asia/Shanghai—says callers can start from a library of 2,000+ voices or a custom-designed voice, then use prompts to fine-tune timbre, pitch, pace, and accent (examples in the post include “add a subtle Southern U.S. accent” and “make it sound like a fire-breathing dragon”). A follow-up in the same thread points to “Remix this voice” in AI Studio, with side-by-side audition of three generated takes, then save for reuse across scripts and multi-speaker scenes. @GoogleDeepMind reposted the same announcement shortly after (2026-11-04T17:12:04Z / 2026-11-05 01:12:04 Asia/Shanghai; https://x.com/GoogleDeepMind/status/2118028516948336696)—useful as a cross-check, not a second product brief. DigIntoAI has not clicked Remix in Studio, has not called the voices API, and has not listened to sample takes; everything below follows those public posts and Google’s developer documentation.

Related DigIntoAI notes: OpenAI Custom Voices—Prompt Design vs Recording Clone, and Where Each Works, Gemini API Pricing (November 2026), Gemini API Now Stores Generate Content Requests by Default on New Paid Projects, and Google Announces Gemini 4 Carbon, Next Frontier Model.

What shipped (feature on existing TTS—not a new model)

Voice Remixing—natural-language edits to an existing voice—lands on the already shipping Gemini 3.8 Flash TTS stack. It is not a new TTS model ID launch and not a replacement for Gemini 3.8 Flash TTS / Flash-Lite TTS themselves.

Google’s Voice remixing documentation frames the flow as:

1. Pick a base_voice (prebuilt studio name, Extended Voice Library ID, or a stored custom voice_...).
2. Call POST /v1beta/voices with type="prompted", store=True, and a relative natural-language input describing the change (age, pitch, energy, timbre, accent, pacing, and similar).
3. Receive a new, independently stored voice_... ID—optionally with a 24 kHz WAV sample_audio preview—without mutating the original base_voice.

That remixed voice_... can then drive synthesis on both gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts. Creation commonly uses model gemini-3.8-flash-tts (the docs list it as the default when omitted). Interactive entry point named in the docs: Google AI Studio generate-speech (https://aistudio.google.com/generate-speech).

The broader speech-generation guide already lists remixed voice_... IDs among the voice options for single-speaker TTS (alongside prebuilt voices, the Extended Voice Library, Voice design, and Voice replication)—so remixing sits inside the existing TTS docs tree, not as a separate model family.

API shape in plain language

Docs examples remix a prebuilt studio voice such as "Puck" with a relative prompt (for example, slightly older, warmer British accent, measured audiobook pacing), print the new voice_... id, optionally save the WAV preview, then synthesize with that id in speech_config.

Key parameters the docs spell out:

Docs also note iterative remixing: a remixed voice_... can itself become the next base_voice, so traits can be refined step by step. Remixing creates a new resource; it does not overwrite the source. Deleting the original base later, per the docs, does not erase voices remixed from it.

Quota notes appear in the same page (as written there): Voice remixing needs a Paid Tier (Tier 1 or higher) Gemini API project; up to 200 stateful stored voices per project (shared across designed, remixed, and stored replicated voices); retention 1 year from last use, with the window resetting when the voice is used for synthesis or as a remix base. DigIntoAI has not verified those limits against a live project bill.

Prompting guidance in the docs favors relative edits (change what matters) rather than restating a full persona; language/accent of the base tend to hold unless the prompt explicitly shifts them; and permanent character traits belong in remix, while turn-level acting cues stay in speech_metadata.style during synthesis.

Two voice counts—separate layers

Public messaging and docs talk about voice inventory at different layers:

• @Google’s post and the September product blog both cite an expansive library of 2,000+ production-ready / preset voices (the post also allows starting from a custom-designed voice).
• Voice remixing docs say base_voice can be any of about 30 prebuilt studio names (examples: "Puck", "Kore", "Zephyr"), or any catalog ID from the Extended Voice Library (GET /v1beta/voices / client.voices.list()), or a project-owned prompted voice_....

Those are not the same set written twice. 2,000+ is the broad library figure in the public announcement and product blog; the ~30 studio names + Extended Voice Library + stored custom IDs is the developer-doc breakdown of what base_voice accepts. DigIntoAI has not enumerated the live catalog.

Timeline: Coming soon on the blog, live in posts + docs

On 2026-09-23, Google’s product blog *Gemini 3.8 text-to-speech says hello* introduced Gemini 3.8 Flash TTS and Flash-Lite TTS and listed Voice remixing under create-and-customize as “Coming soon”—pick a library voice and fine-tune timbre, pitch, pace, and accent with prompts. Checking that page again, the Coming soon wording is still visible, and dateModified still sits on 2026-09-23 (no past-tense “now live” rewrite on that blog line).

Whether remixing is available now follows the 2026-11-04 / 2026-11-05 Asia/Shanghai @Google post plus the current Voice remixing developer docs—not the unchanged Coming soon sentence on the September blog. The blog remains useful for the earlier timeline and for the shared 2,000+ library language; it should not be read as “Google never launched remixing,” and it should not be rewritten as if that September page already announced GA.

What DigIntoAI has not done

• No Google AI Studio login to click Remix this voice or compare three takes
• No POST /v1beta/voices call from DigIntoAI
• No listening evaluation of remixed sample audio, accents, or multi-speaker reuse
• No billing or quota measurement for remix create vs TTS synthesize
• No claim that the September product blog was updated to remove Coming soon

Product posts and docs can change; Google’s public posts and the Voice remixing / speech-generation documentation remain the primary sources.

Quick facts

1. What. Voice Remixing—natural-language edits to an existing voice → new voice_... ID.
2. Where. Google AI Studio + Gemini API (per @Google); Studio generate-speech linked from docs.
3. When announced live. @Google 2026-11-05 01:06:16 Asia/Shanghai (2026-11-04T17:06:16Z); DeepMind repost minutes later.
4. Models that use the remixed ID. gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts (not a new model family).
5. API. POST /v1beta/voices, type="prompted", base_voice + relative input, store=True.
6. Library language. Public post / Sept blog: 2,000+; docs: ~30 studio names + Extended Voice Library + stored customs—layers stay separate.
7. Sept 23 blog. Still shows Voice remixing Coming soon; availability follows posts + current docs.
8. Related. OpenAI custom voices (prompt vs recording), Gemini API pricing, Generate Content default store, Gemini 4 Carbon.

In short

Google says Voice Remixing for Gemini 3.8 Flash TTS is live in AI Studio and the Gemini API, with plain-language control over timbre, pitch, pace, and accent from a 2,000+ (or custom) starting point. Developer docs spell out POST /v1beta/voices → a new stored voice_... for gemini-3.8-flash-tts / gemini-3.8-flash-lite-tts, without overwriting the base. That is a feature on the existing 3.8 TTS models, not a fresh model launch. The 2026-09-23 product blog can still show Coming soon; the live signal is the November public posts and current docs. DigIntoAI hasn't tested Studio, the API, listening quality, or billing—details above follow those public sources.